Satellite-Based Prediction of Arctic Sea Ice Concentration Using a Deep Neural Network with Multi-Model Ensemble

• Published in: Remote sensing (Basel, Switzerland) Vol. 11; no. 1; p. 19
• Main Authors: Kim, Jiwon, Kim, Kwangjin, Cho, Jaeil, Kang, Yong, Yoon, Hong-Joo, Lee, Yang-Won
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01-01-2019
• Subjects: Accuracy; Algorithms; Artificial neural networks; Bayesian analysis; Bayesian model averaging; Climate change; Climate models; Computer centers; Correlation coefficient; Correlation coefficients; deep neural network; Forecasting; General circulation models; Ice; Ice environments; Linear equations; Linearity; Machine learning; Microwave sensors; Neural networks; Nonlinearity; regional climate model; Remote sensing; Sea ice; sea ice concentration; Solar energy; Spatial data; Spatial resolution; Spatial variations; Structure-function relationships; Temperature; Uncertainty; Variables; Arctic Ocean; Barents Sea; United States > US; Hudson Bay; Arctic region; Novaya Zemlya
• DOI: 10.3390/rs11010019
• EISSN: 2072-4292

Warming of the Arctic leads to a decrease in sea ice, and the decrease of sea ice, in turn, results in warming of the Arctic again. Several microwave sensors have provided continuously updated sea ice data for over 30 years. Many studies have been conducted to investigate the relationships between the satellite-derived sea ice concentration (SIC) of the Arctic and climatic factors associated with the accelerated warming. However, linear equations using the general circulation model (GCM) data, with low spatial resolution, cannot sufficiently cope with the problem of complexity or non-linearity. Time-series techniques are effective for one-step-ahead forecasting, but are not appropriate for future prediction for about ten or twenty years because of increasing uncertainty when forecasting multiple steps ahead. This paper describes a new approach to near-future prediction of Arctic SIC by employing a deep learning method with multi-model ensemble. We used the regional climate model (RCM) data provided in higher resolution, instead of GCM. The RCM ensemble was produced by Bayesian model averaging (BMA) to minimize the uncertainty which can arise from a single RCM. The accuracies of RCM variables were much improved by the BMA2 method, which took into consideration temporal and spatial variations to minimize the uncertainty of individual RCMs. A deep neural network (DNN) method was used to deal with the non-linear relationships between SIC and climate variables, and to provide a near-future prediction for the forthcoming 10 to 20 years. We adjusted the DNN model for optimized SIC prediction by adopting best-fitted layer structure, loss function, optimizer algorithm, and activation function. The accuracy was much improved when the DNN model was combined with BMA2 ensemble, showing the correlation coefficient of 0.888. This study provides a viable option for monitoring Arctic sea ice change of the near future.